Advanced Ensemble Learning Classifiers for Power Transformer Fault Diagnosis with Imbalanced DGA Data

Abstract

Power transformers play a vital role in the reliability and functionality of power systems. In recent years, the growing complexity of power transformer fault diagnosis has driven the widespread adoption of machine learning classifiers. This paper explores the application of tree-based machine learning algorithms, including Decision Tree, AdaBoost, and CatBoost, to diagnose power transformer faults using Dissolved Gas Analysis (DGA) data, normalized via standard scaling. Specifically, the Duval Pentagon Method (DPM) is utilized as a fault classification tool. One of the key challenges addressed in this study is the inherent imbalance in DGA datasets, which is tackled effectively through the application of robust tree-based models. A comprehensive comparative analysis of these algorithms is performed, with their performance evaluated based on metrics such as accuracy, precision, F1-score, area under the curve (AUC), receiver operating characteristic (ROC) curve, and confusion matrix. The simulation results, obtained using Python, reveal that the CatBoost algorithm significantly outperforms its counterparts, achieving an impressive accuracy of 96.78%, F1-score of 96.78%, precision of 96.81%, and AUC of 99.79%. These findings highlight the effectiveness of CatBoost in handling imbalanced data and providing precise fault diagnosis, contributing to improved decision-making processes in power system maintenance and reliability management.